Author: Carly Eckert

Coauthor(s): Nick Mark, MD Muhammad Ahmad, PhD Greg McKelvey, MD MPH

Status: Project Concept

Solid organ transplantation is an important lifesaving intervention, which is unfortunately limited by a dearth of organs. Currently, more than 30,000 organs are transplanted in the US each year, and the number of transplants performed is consistently increasing. Despite this relative increase in transplants, organ shortage continues to be a significant problem. There are over 100,000 patients awaiting transplant in the US, the majority awaiting kidneys. The median wait time to first kidney transplant is 3.6 years, and over 5,000 people die awaiting kidney transplant. As many as 20% of potentially viable donor kidneys are discarded, partially due to inefficiencies in the logistics of procurement and allocation. Models to predict organ availability could thus benefit hundreds of patients annually.

Currently, organ availability for transplantation is viewed as a purely stochastic event. Transplant centers and teams discuss heuristics like “there are usually transplants after holiday weekends” but assumptions are based on anecdotes, not data. Yet, there are multiple decision nodes between donor death and organ transplant that could be used to develop predictive models. The United Network for Organ Sharing (UNOS) maintains a dataset of hundreds of variables on every transplant-related event in the US since 1988, a database of millions of encounters for kidneys alone. This dataset includes donor and recipient variables, transplant-related variables, waiting list data, and patient outcomes and follow-up. Date, time, and location of events are also included in the dataset. To the best of our knowledge, machine learning and predictive analytics have never been applied to these data.

We propose using machine learning algorithms to help organ procurement organizations benefit from knowledge around patterns of organ availability. Predictive models could be used to optimize organ procurement and improve outcomes after transplant. Planning and dispatching teams for organ procurement could be more easily coordinated with the help of predictive models – saving these organizations valuable resources and potentially reducing organ ischemic time. Patterns around organ availability could be related to events leading to organ donors (car crashes, strokes, violent crime) as well as geographic, seasonal, weekly, or daily patterns could be used to predict “surges” in the availability of transplanted organs. Additionally, transplant centers face significant issues with planning for organ transplants, issues that can be mitigated by machine learning. Operating rooms and appropriate staff need to be available with little to no warning as organ transplant viability (i.e. graft survival) is highly dependent on time to transplant. However, keeping ORs and transplant staff “on-call” can be expensive and a significant opportunity cost for hospitals that could otherwise direct these resources towards other patients. Additionally, transplant physicians and surgeons often have their schedules, clinics, and nights disrupted by transplant call. While most transplant physicians consider this part of the job, they could certainly benefit from models that better predict when they are needed.

Using the complete UNOS database, we are developing a suite of ML predictive models related to organ transplants, leveraging the rich UNOS data, to better serve those do this valuable work.